Korva is a scroll of the internet’s best ideas.
Software has a shadow economy
Anyone can rewrite an application away. People have vibe coded their own Salesforce and boasted about it. What that misses is the work downstream: people with no Salesforce seat license carry out processes and tasks that result from the information coming out of that database.
That creates a shadow economy inside the customer, a derivative effect of the system that was put in place. The system has usually been running for years and executes with deterministic outcomes, otherwise it wouldn't still be there. Many people who never pay for it depend on its decisions as a step in their workflow.
A good CIO asked about ripping it out will admit not knowing what would happen, so companies try little pieces first. Where a piece is materially better than what came before, it can be replaced. Customer support and contact centers are an example, because an agent-driven tier one and tier two support experience is better.
Why do old systems keep running for decades?
Replacing a working system risks breaking every process that quietly depends on it. Many banks still run their core ledgers on code written in COBOL in the 1960s and 1970s, because those systems process transactions correctly and the cost of an error is enormous. Each year a system runs adds integrations, reports, and habits built around it. That accumulated web, more than the code itself, is what makes migration projects slow and expensive.
Why do deterministic outcomes matter so much?
A deterministic system gives the same output every time it receives the same input. That lets auditors trace a result back to its cause and lets other systems rely on it without checking. Large language models generate text by sampling from probabilities, so the same prompt can return different answers on different runs. Engineers often add validation layers around model output for exactly that reason.
How do seat licenses hide who depends on software?
A seat license charges for each named user who logs in, so the vendor's revenue counts people with access rather than people affected by the system. Reports, approvals, exports, and data feeds from the system reach many employees who never open it. Those downstream consumers do not appear on the license list. A map of who depends on a system's outputs has to be built separately from the list of who pays for it.
More from Joys of Compounding
